Robust parallel decision-making in neural circuits with nonlinear inhibition

• DOI: 10.1073/pnas.1917551117
• Publication Date: 2020-10-03T00:26:17Z
• Authors: Birgit Kriener; Rishidev Chaudhuri; Ila R. Fiete
• Keywords: speed–accuracy trade-off; Neurons; 0301 basic medicine; 0303 health sciences; Neural Networks; speed-accuracy trade-off; Decision Making; Models, Neurological; Neurosciences; Biological Sciences; optimal decision-making; Computer; 03 medical and health sciences; Nonlinear Dynamics; Models; Neurological; Neural Networks, Computer; noisy computation; Nerve Net; neural circuits

An elemental computation in the brain is to identify best a set of options and report its value. It required for inference, decision-making, optimization, action selection, consensus, foraging. Neural computing considered powerful because parallelism; however, it unclear whether neurons can perform this max-finding operation way that improves upon prohibitively slow optimal serial (which takes [Formula: see text] time N noisy candidate options) by factor N, benchmark parallel computation. Biologically plausible architectures task are winner-take-all (WTA) networks, where individual inhibit each other so only those with largest input remain active. We show conventional WTA networks fail parallelism and, worse, presence noise, altogether produce winner when large. introduce nWTA network, which equipped second nonlinearity prevents weakly active from contributing inhibition. Without parameter fine-tuning or rescaling as varies, network achieves benchmark. The reproduces experimentally observed phenomena like Hick's law without needing an additional readout stage adaptive N-dependent thresholds. Our work bridges scales linking cellular nonlinearities circuit-level establishes distributed saturating possible noisy, finite-memory neurons, shows may be symptom near-optimal decision-making input.